The Agency: hire 279 AI specialists

1. 01

   It's a folder of job descriptions

   279 Markdown files, each one an AI specialist. No code runs: your AI tool reads the file when that specialist is called in.

   How
2. 02

   Ask for a specialist by name

   Instead of a generic assistant, you say which agent you want. In Claude Code it works in its own context and reports back.

   How
3. 03

   You can try one with no setup

   Copy one agent file from GitHub, paste it into any AI chat, then add your ask. Install only when you like it.

   How
4. 04

   Every file has the same shape

   Frontmatter, identity, mission, critical rules, deliverables, workflow and metrics. The critical rules are the part to edit for your own stack.

   How
5. 05

   A team beats one agent

   Name several agents in one prompt, in order. Each hands its work to the next, and a checker goes last.

   How
6. 06

   NEXUS runs them like a company

   Seven phases from Discover to Operate, with a sign-off gate between each. Start with the smallest mode, Micro: 5 to 10 agents.

   How
7. 07

   Built, tested, three strikes

   Each task is built, then tested with screenshots: PASS or FAIL. After three FAILs the Orchestrator reassigns, splits, rethinks, accepts or defers it.

   How
8. 08

   Two agents are built to say no

   Evidence Collector assumes there are 3 to 5 issues and wants proof. Reality Checker's default answer is NEEDS WORK.

   How
9. 09

   Install a few, not all 279

   Every installed agent takes up room in each session. Pick a division or single agents with the app or the install script.

   How

Agent (persona)

One Markdown file that describes an AI specialist: its role, personality, rules, example work and workflow. Your AI tool reads it when that specialist is called in.

Markdown

Plain text with light formatting, saved as a .md file. Every agent in the repo is one of these. No code runs.

Repo

Short for repository: a project's folder of files on GitHub. This one is msitarzewski/agency-agents.

Clone

Download a copy of a repo to your computer with git clone. The install scripts run from inside that copy.

Claude Code

An AI coding tool, and the main one this page installs into. Agent files go in ~/.claude/agents/, so every project gets them.

Context window

The working memory of one AI session. Every installed agent's name and description take up room in it, which is why fewer agents is better.

Token

The unit AI models count text in. FakeNerd's rough rule here: one token for every four characters.

Division

A department of agents, like Engineering, Marketing or Testing. There are 18. The installer can add a whole division at once with --division.

Tools (Bash, Write)

The actions an agent is allowed to take, like running commands (Bash) or writing files (Write). An agent file without a tools: line gets every tool your session has.

Agents Orchestrator

The agent that runs the others in NEXUS. It hands out tasks, runs the Dev↔QA loop and decides what happens after three failed tries.

QA

Quality assurance: testing work before it ships. Here the QA agent is Evidence Collector, which tests each task with screenshots and says PASS or FAIL.

MVP

Minimum viable product: the first small version of a product you can launch. NEXUS-Sprint mode is sized for a feature or an MVP.

• 279agent files
• 18divisions
• ~75klines of persona
• 16tools it installs into

1. 01

   Copy and paste

   Any AI chat. No setup.

   1. Open an agent file on GitHub, like the SEO Specialist.
   2. Copy the whole file.
   3. Paste it into ChatGPT, Claude or Gemini, then add your ask.

   Your message

   [paste the agent file]
   
   Now audit fakenerd.ai's homepage for SEO.

2. 02

   Install as a subagent

   Claude Code, Codex, Cursor and 13 more.

   1. Install with the Agency Agents app, or the script below.
   2. The files land in ~/.claude/agents/, so every project gets them.
   3. Ask for one by name. It works in its own context and reports back.

   Type this

   Use the SEO Specialist agent to audit the /pstack page.

3. 03

   Run a team

   When one job needs several skills.

   1. Name several agents in one prompt, in order.
   2. Each hands its work to the next.
   3. End with a checker, so nothing ships on vibes.

   Type this

   Activate UX Researcher to find usability issues on the homepage.
   UX Architect designs fixes. Frontend Developer builds them.
   Evidence Collector verifies with screenshots.

./scripts/install.sh --tool claude-code \
  --agent seo-specialist,ai-citation-strategist,content-creator,brand-guardian,reality-checker

1. ---
   name: Frontend Developer
   description: Expert frontend developer specializing in modern web…
   color: cyan
   emoji: 🖥️
   vibe: Builds responsive, accessible web apps with pixel-perfect precision.
   ---

   01Frontmatter

   What the tool reads. name + description are how Claude Code decides when to use it. vibe and emoji are for people and the app.

2. ## 🧠 Your Identity & Memory
   - Role: Modern web application and UI implementation specialist
   - Personality: Detail-oriented, performance-focused…

   02Identity & memory

   A role and a temperament. This is the “personality” the repo is known for.

3. ## 🎯 Your Core Mission
   ### Create Modern Web Applications
   - Default requirement: accessibility and mobile-first

   03Core mission

   What it's for, in a few headed lists. Usually one non-negotiable default.

4. ## 🚨 Critical Rules You Must Follow
   - Follow WCAG 2.1 AA guidelines
   - Implement Core Web Vitals optimization from the start

   04Critical rules

   The guardrails. The most useful part to edit for your own stack.

5. ## 📋 Your Technical Deliverables
   ```tsx
   export const DataTable = memo<DataTableProps>(…)

   05Deliverables

   Worked examples: code, templates, report formats. Shows the model what “done” looks like.

6. ## 🔄 Your Workflow Process
   ## 🎯 Your Success Metrics
   - Page load times under 3 seconds on 3G

   06Workflow + metrics

   Steps it should follow and how to tell if it worked.

Read the full file. In Claude Code these are subagents: each one runs in its own context window, so a long review doesn't clutter your main chat.

Browse the roster at fakenerd.ai/agency-agents#roster or in the repo.

Activate Backend Architect to investigate and fix [BUG DESCRIPTION].
After fix, activate API Tester to verify the fix.
Then activate Evidence Collector to confirm no visual regressions.

Want to see a big one run? In the Nexus Spatial Discovery example, eight agents worked in parallel and produced a market check, architecture, brand, go-to-market, support plan, UX research and a 35-week plan in one session. More examples.

1. Phase 0

   Discover

   Market, users, competitors, compliance.

   Gate: Executive Summary Generator

2. Phase 1

   Strategize

   Architecture, brand, backlog, budget.

   Gate: Studio Producer + Reality Checker

3. Phase 2

   Scaffold

   CI/CD, design system, app skeleton.

   Gate: DevOps Automator + Evidence Collector

4. Phase 3

   Build

   Task by task, each through the Dev↔QA loop.

   Gate: Agents Orchestrator

5. Phase 4

   Harden

   Performance, API, compliance, full journeys.

   Gate: Reality Checker, alone

6. Phase 5

   Launch

   Content, channels, deploy, growth.

   Gate: Studio Producer + Analytics Reporter

7. Phase 6

   Operate

   Monitor, support, report, improve.

   Gate: Ongoing

• Quality gates

  No phase moves on without evidence-based sign-off.

• Dev↔QA loop

  Every task is built, then tested. Fail goes back with notes. Three strikes, escalate.

• Handoffs

  Structured context goes with every task. No agent starts cold.

• Evidence over claims

  Screenshots, test results and data. Not “looks great, A+”.

Every hand-off uses a template: from, to, phase, task, priority, context, what's wanted and how it'll be judged.

1. A developer agent builds one task.
2. Evidence Collector tests it with screenshots: PASS or FAIL.
3. FAIL goes back with specific issues and files. Fix only, no new features.
4. After three FAILs, escalate: reassign, split, rethink, accept or defer.

1. 01Context

   Don't install all 279

   Claude Code shows the model every installed agent's name and description so it can pick one. All 279 is about 67,500 characters, roughly 17,000 tokens, in every session. Install two or three divisions. Heads-up: --division marketing also brings 14 China-platform agents (WeChat, Douyin, Weibo…); use --agent to pick.

2. 02Permissions

   Most agents get every tool

   Only 17 of 279 files set a tools: line. The rest inherit every tool your session has, including Bash and Write. Add tools: to the ones that only need to read or write docs.

3. 03Fit

   Some agents expect a stack you don't have

   Reality Checker and Evidence Collector run a qa-playwright-capture.sh the repo doesn't ship, and look in Laravel folders. Swap in your own screenshot step and paths.

4. 04Names

   The Orchestrator uses old nicknames

   Its file asks for “EvidenceQA” and “ArchitectUX”. The installed agents are Evidence Collector and UX Architect. Rename them in your copy so hand-offs land.

5. 05Claims

   The NEXUS numbers have no source

   The executive brief says 73% of handoffs fail, 40–60% faster, 80% fewer defects. No study, no method. Read them as a pitch, not data.

6. 06Use

   The best part is the skeptics

   Evidence Collector (find 3–5 issues, show proof) and Reality Checker (default: NEEDS WORK) fix the most common AI failure: grading its own work A+. Steal those two first.

• msitarzewski/agency-agents by Michael Sitarzewski and contributors, MIT license. Agent names, emoji, “vibe” lines, teams and NEXUS structure are theirs; read at commit 053ddbb.
• NEXUS docs: quick-start, full strategy, executive brief, hand-off templates.
• agencyagents.app, the desktop installer.
• Counts, the token estimate, the four prompts marked as ours and the take section are FakeNerd's.